Harness learning: frozen-weight agents improve by rewriting their own harness code, 32%→62%
mark_k · x · 2026-10-12
A new paper introduces harness learning: training a model to revise another agent's harness — the code controlling its model calls, tools and information flow. It reads execution feedback, changes the workflow, and retries; once trained, it works on new tasks without further weight updates.
- Across 21 unseen reasoning task families, average scores rose from 32% with the untrained revision model to 62% after RL.
- The trained 4B model produced better revisions on average than its 35B teacher.
- In multi-hop QA, improvements continued over repeated revision rounds and transferred to unseen benchmarks.
The transferability of workflow-improvement ability is the key finding — a concrete reason to pursue coding agents that learn from failed runs and adapt how they search, test and use tools.
Related event: Harness Learning: Agents Adapt at Test Time Without Weight Updates(3 posts)→
More from coding & agent
- Dev builds Claude-coded 'daily wearable' firmware packing 27 games and 16 utilities — ezshine · 2026-10-12
- MCP memory app keeps persistent context across Claude, ChatGPT and Codex — Jolly-Scientist-2484 · 2026-10-12
- HANDBOOK.md benchmark: top agents pass just 36.2% of 65 policy-heavy agentic tasks — bibryam · 2026-10-12
- LangChain's Harrison Chase backs the bet that AI feeds will be the form factor of 2027 — hwchase17 · 2026-10-12
- AWS open-sources Strands Decider 2B, a laptop-runnable decision model with no text generation — dl_weekly · 2026-10-12
- aiprox-mcp lets you discover, hire and pay autonomous AI agents across Lightning, Solana and x402 rails — modelcontextprotocol · 2026-10-12